Ivy-Fake
Multimodal dataset bringing together annotated images and videos for the detection of synthetic artifacts related to generation AI, with detailed explanations.
Over 150,000 training samples (images and videos), 18,700 evaluation samples, fine annotations
Apache 2.0
Description
Ivy-Fake is a unique multimodal dataset with over 150,000 examples of synthetic and real images and videos, annotated with binary labels (real/false) and detailed explanations. It aims to enable the automated and explainable detection of AI-generated content (AIGC), by analyzing both spatial and temporal artifacts.
What is this dataset for?
- Train vision and language models to detect multimodal synthetic content
- Develop explainability systems to understand model decisions
- Improving the robustness of detection systems in the face of AI-generated content
Can it be enriched or improved?
This dataset can be enriched by adding new synthetic content classes, or by creating additional annotations that further detail the types of artifacts observed. Integrating data from other multimodal sources can also increase its diversity and relevance.
🔎 In summary
🧠 Recommended for
- Vision-language researchers
- AI security developers
- Explainability projects
🔧 Compatible tools
- PyTorch
- TensorFlow
- OpenCV
- Multimodal frameworks
💡 Tip
Use explanatory annotations to train models that generate understandable human justifications.
Frequently Asked Questions
Does this dataset include videos and images?
Yes, it contains over 150,000 multi-modal examples including images and videos.
What is the main purpose of Ivy-Fake?
It is mainly used to train and evaluate models capable of detecting and explaining artificial synthesis in the media.
What license protects this dataset?
It is licensed under Apache 2.0, allowing commercial and free use.
